Use cupy....The installation is a bit complex. Need to be careful about the environment variables.
initialize/update the weight by numpy
Reference to normal initializer
Its implementation is different from PyTorch.
要使用 GPU 的话， 需要 pygpu 库， 好像只能用 conda 安装， pip 是找不到这个库的.
We need pygpu library to utilize GPU, so it seems that only conda support pygpu, I can't find the library by pip.
tensorboardX need to use pip to install.
this man is lazy, nothing left